- AA Index (v5.0) 20.1
- Vals Index (v2.2) 46.5
- FrontierCode v2.0 (stirrup v0.2.1) 15.1
- DeepSWE v1.1 (mini-swe-agent v2.4.6) 48.2
- Terminal Bench v4.0 (mini-swe-agent v2.4.6) 5.6
- Terminal Bench v2.1 (Inkling 2 Mini in pi v0.83.0) 71.6
- SWEBench Pro v2 (Hard), mini-swe-agent v2.4.6 77.9
- SWEBench Multilingual, bash only 73.7
- SWEBench Verified, bash only 79.2
- Vibe Code Bench v1.1 (openhands, Vals) 77.9
- PMPP-Hard (AUC, 0 to 10h), pi v0.83.0 455.8
- BrowseComp with compaction 79
- DeepSearchQA with compaction 85.3
- MCP Atlas (Public) 82.7
- AA-Briefcase v1.1 1263
- GDPVal-AA v2.1 1319
- JobBench 49.6
- AutomationBench-AA v1.0.6 41.8
- EnterpriseOps-Gym 42
- Vals Finance Index 50.1
- AA-AnalystAgent 46.3
- Diligence Bench (best harness) 48.6
- GDP.pdf 16.8
- PerceptionBench 57.5
- MMMU Pro (Standard 10) 74.9
- OmniDocBench v1.6 92.8
- ScreenSpot Pro (original, 87.7 with Python) 76.5
- Chartography (original, 72.5 with Python) 22.1
- ZeroBench (main, original, 34.0 with Python) 17
- MuirBench 77.5
- BVB 68.4
- HLE-Diamond 15.2
- CritPt v2 (Full, 17.0 on Hard) 31.4
- Apex 2025 34.4
- ARC-AGI v1 84.5
- ARC-AGI v2 49.3
- LongBench v2 66.9
- GraphWalks BFS 256K 69.7
- MRCR v2 8-needle 256K to 512K 46.4
- AA-Omniscience v1.1 score (non-hallucination rate 63.4%) -8.6
- ComplexConstraints v1.1 (openhands v1.45.0) 33.4
- FORTRESS (adversarial 89.5%, benign 96.4%), Public
- StrongREJECT (none) 97.8
- Misbehavior (Internal), lower is better 9.8
- Covert misbehavior (Internal), lower is better 8.3
- Destructive actions (Internal), lower is better 10.2
Inkling 2 Mini
Inkling 2 Mini.
Inkling 2 Mini
Open weights Small
- GraphWalks 69.7% #20 /29
- AA-LCR 1.1 80% #4 /7
- StrongREJECT 97.8% #5 /20
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AA-LCR 1.1 | 80% | #4 of 7 lab numbers | Lab model card | |
| GraphWalks | 69.7% | #20 of 29 lab numbers | Lab model card | |
| AA-Omniscience v1.1 | -8.6 | #2 of 2 lab numbers | Lab model card | |
| ComplexConstraints | 33.4% | Not ranked | Lab model card | |
| HLE-Diamond | 15.2% | Not ranked | Lab model card | |
| Humanity's Last Exam | 30.7% | Not ranked | Lab model card | |
| LongBench v2 | 66.9% | Not ranked | Lab model card | |
| OpenAI MRCR v2 | 46.4% | Not ranked | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ARC-AGI | 84.5% | Not ranked | Lab model card | |
| ARC-AGI-2 | 49.3% | Not ranked | Lab model card | |
| CritPt v2 | 31.4% | Not ranked | Lab model card | |
| GPQA Diamond | 86.6% | Not ranked | Lab model card | |
| MathArena (Apex, IMO, USAMO) | 34.4% | Not ranked | Lab model card | |
| MysteryMechanism | 23% | Not ranked | Lab model card | |
| ProofBench v1.1 | 18% | Not ranked | Lab model card | |
| Terminal-Bench-Science | 1.40% | Not ranked | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Code2Video-AA | 870 | #2 of 2 lab numbers | Lab model card | |
| Code Migration v1.1 | 23.2% | Not ranked | Lab model card | |
| DeepSWE | 48.2% | Not ranked | Lab model card | |
| FrontierCode 2.0 | 15.1% | Not ranked | Lab model card | |
| PMPP-Hard | 455.8 | Not ranked | Lab model card | |
| SWE-bench Multilingual | 73.7% | Not ranked | Lab model card | |
| SWE-bench Multimodal v2 | 37.9% | Not ranked | Lab model card | |
| SWE-Bench Pro V2 Hard | 77.9% | Not ranked | Lab model card | |
| SWE-bench Verified | 79.2% | Not ranked | Lab model card | |
| Vibe Code Bench | 77.9% | Not ranked | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AutomationBench | 41.8% | Not ranked | Lab model card | |
| BrowseComp | 79% | Not ranked | Lab model card | |
| DeepSearchQA | 85.3% | Not ranked | Lab model card | |
| EnterpriseOps-Gym | 42% | Not ranked | Lab model card | |
| MCP Atlas | 82.7% | Not ranked | Lab model card | |
| ScreenSpot-Pro | 76.5% | Not ranked | Lab model card | |
| Terminal-Bench 2.1 | 71.6% | Not ranked | Lab model card | |
| Terminal-Bench 4.0 | 5.60% | Not ranked | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GDPval-AA v2.1 | 1,319 | #10 of 11 lab numbers | Lab model card | |
| AA-Briefcase v1.1 | 1,263 | #10 of 10 lab numbers | Lab model card | |
| AA-AnalystAgent | 46.3% | #2 of 2 lab numbers | Lab model card | |
| DiligenceBench | 48.6% | #2 of 2 lab numbers | Lab model card | |
| Excel Modeling Benchmark v1.1 | 66% | Not ranked | Lab model card | |
| JobBench | 49.6% | Not ranked | Lab model card | |
| Tax Agent Bench | 14.1% | Not ranked | Lab model card | |
| Vals Finance Agent v2.1 | 56.5% | Not ranked | Lab model card | |
| Vals Finance Index | 50.1% | Not ranked | Lab model card | |
| Vals Index v2.2 | 46.5% | Not ranked | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BabyVision | 53.4% | Not ranked | Lab model card | |
| BVB | 68.4 | Not ranked | Lab model card | |
| Chartography | 22.1% | Not ranked | Lab model card | |
| CharXiv | 80.6% | Not ranked | Lab model card | |
| GDP.pdf | 16.8% | Not ranked | Lab model card | |
| MATH-Vision (MathVision) | 89% | Not ranked | Lab model card | |
| MMMU-Pro | 74.9% | Not ranked | Lab model card | |
| MuirBench | 77.5% | Not ranked | Lab model card | |
| OmniDocBench v1.6 | 92.8 | Not ranked | Lab model card | |
| PerceptionBench (Moonshot) | 57.5% | Not ranked | Lab model card | |
| ZeroBench | 17% | Not ranked | Lab model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| StrongREJECT | 97.8% | #5 of 20 lab numbers | Lab model card | |
| Covert misbehavior (Thinking Machines) | 8.30% | #1 of 2 lab numbers | Lab model card | |
| Misbehavior (Thinking Machines) | 9.80% | #1 of 2 lab numbers | Lab model card | |
| Destructive actions (Thinking Machines) | 10.2% | #2 of 2 lab numbers | Lab model card | |
| HoneyBench v0.2-alpha | 56.7 | #2 of 2 lab numbers | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5 | 20.1 | Not ranked | Lab model card | |
| SAGE v1.1 (Vals) | 85.9% | Not ranked | Lab model card |
Its lab's numbers put it at 148.8, number 111 of 412
The Self-Reported Index fits one score to the 28 capability results Thinking Machines reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 146.5 to 151.1. How we compute it.
- MATH-Vision (MathVision) 89
- GPQA Diamond 86.6
- DeepSearchQA 85.3
- ARC-AGI 84.5
- MCP Atlas 82.7
- CharXiv 80.6
- SWE-bench Verified 79.2
- BrowseComp 79
- MuirBench 77.5
- ScreenSpot-Pro 76.5
- MMMU-Pro 74.9
- SWE-bench Multilingual 73.7
- Terminal-Bench 2.1 71.6
- GraphWalks 69.7
- LongBench v2 66.9
- BabyVision 53.4
- JobBench 49.6
- ARC-AGI-2 49.3
- DeepSWE 48.2
- OpenAI MRCR v2 46.4
- AutomationBench 41.8
- MathArena (Apex, IMO, USAMO) 34.4
- Humanity's Last Exam 30.7
- Chartography 22.1
- ZeroBench 17
- GDP.pdf 16.8
- Terminal-Bench 4.0 5.6
- Terminal-Bench-Science 1.4
Who measured it?
No independent run has it yet. All 62 of its scores are Thinking Machines's own numbers or claims Epoch AI collected.
- A claim only 62
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 17 of them aren't in the launch post.
Capabilities
- Artificial Analysis Intelligence Index v5 20.1
- Vals Index v2.2 46.5
- FrontierCode 2.0 15.1
- DeepSWE 48.2
- Terminal-Bench 4.0 5.6
- Terminal-Bench 2.1 71.6
- SWE-Bench Pro V2 Hard 77.9
- SWE-bench Multilingual 73.7
- SWE-bench Verified 79.2
- Vibe Code Bench 77.9
- PMPP-Hard 455.8
- BrowseComp 79
- DeepSearchQA 85.3
- MCP Atlas 82.7
- AA-Briefcase v1.1 1263
- GDPval-AA v2.1 1319
- JobBench 49.6
- AutomationBench 41.8
- EnterpriseOps-Gym 42
- Vals Finance Index 50.1
- AA-AnalystAgent 46.3
- DiligenceBench 48.6
- GDP.pdf 16.8
- PerceptionBench (Moonshot) 57.5
- MMMU-Pro 74.9
- OmniDocBench v1.6 92.8
- ScreenSpot-Pro 76.5
- Chartography 22.1
- ZeroBench 17
- MuirBench 77.5
- BVB 68.4
- HLE-Diamond 15.2
- CritPt v2 31.4
- MathArena (Apex, IMO, USAMO) 34.4
- ARC-AGI 84.5
- ARC-AGI-2 49.3
- LongBench v2 66.9
- GraphWalks 69.7
- OpenAI MRCR v2 46.4
- ComplexConstraints 33.4
- SWE-bench Multimodal v2 37.9
- Code Migration v1.1 23.2
- Terminal-Bench-Science 1.4
- Vals Finance Agent v2.1 56.5
- Excel Modeling Benchmark v1.1 66
- Tax Agent Bench 14.1
- BabyVision 53.4
- CharXiv 80.6
- SAGE v1.1 (Vals) 85.9
- MATH-Vision (MathVision) 89
- Code2Video-AA 870
- Humanity's Last Exam 30.7
- GPQA Diamond 86.6
- MysteryMechanism 23
- ProofBench v1.1 18
- AA-LCR 1.1 80
Hallucination
- AA-Omniscience v1.1 -8.6